Classifying Marine Mammals using Convolutional Neural Networks

The problem

Marine biologists track individual whales and dolphins by the shapes of their flukes, fins, and scars. Matching photos by hand is slow and gets harder as the datasets grow. The Happywhale dataset has over 51,000 photos of 15,587 individual animals from more than 30 species.

What I built

• An EfficientNetB5 backbone with an ArcFace loss, which pushes different animals further apart in the feature space.
• Augmentations (random crops, grayscale, color changes) so the model holds up across different cameras and lighting.
• A k-nearest-neighbors step at inference to decide when a photo shows an animal it has never seen.
• Result: a Mean Average Precision at 5 of 0.88. In Kaggle's Happywhale competition, my model placed 100th out of about 1,600 teams.

How it happened

I did this independently between March 2024 and September 2025, mentored by Arpit Jasapara, a UCLA master's student in AI and machine learning. Getting through peer review took about three months.